Sakana AI unveiled an open-source algorithm on Tuesday designed to enable various artificial intelligence (AI) models to work together on intricate challenges. The algorithm, referred to as Adaptive Branching Monte Carlo Tree Search (AB-MCTS), serves as an inference-time scaling tool that introduces an additional dimension to the existing framework of AI models. This innovation allows the system to assess whether a longer reasoning process or a broader exploration approach is appropriate when confronting a new problem, as well as determining the most suitable AI model for the task. In situations where a problem proves to be particularly complex, the algorithm can activate multiple AI models simultaneously.
Sakana AI Releases Algorithm That Promotes Collaborative Thinking Among AI Models
The Tokyo-based AI firm announced the release of its algorithm in a tweet. The company emphasized that this new algorithm fosters an environment for collective intelligence in AI, allowing leading models such as Gemini 2.5 Pro, o4-mini, and DeepSeek-R1 to work collaboratively.
Sakana AI has been focused on a challenging issue in the realm of artificial intelligence: integrating the distinct strengths of individual AI models while mitigating their unique biases to enhance overall performance. This research has been ongoing for several years, and in 2024, the company published findings related to “evolutionary model merging.”
Building on this research, Sakana AI has launched an algorithm that enables AI models to conduct computational tasks within specific parameters, allowing them to produce multiple outputs to consider different viewpoints and deploy multiple models suited to the task for improved performance.
Researchers involved in the project successfully assessed the capabilities of the AB-MCTS system using the ARC-AGI-2 benchmark. This system, combining o4-mini, Gemini-2.5-Pro, and R1-0528, outperformed the individual models in performance metrics. According to Sakana AI, while the o4-mini model independently solved 23 percent of the problems, its success rate increased to 27.5 percent when integrated into the AB-MCTS framework.
Sakana AI has made the TreeQuest algorithm available on its GitHub repository and has also published its ARC-AGI experiment findings separately. In addition, details from the study have been published in a paper on arXiv.
